By Venkatesh Pallipadi

Introduction

The Pentium® M processor supports Enhanced Intel SpeedStep® Technology as an advanced means of enabling very high performance while also meeting the power-conservation needs of mobile systems. Conventional Intel SpeedStep Technology switches both voltage and frequency in tandem between high and low levels in response to processor load. Enhanced Intel SpeedStep Technology builds upon that architecture using design strategies that include the following:

• Separation between Voltage and Frequency Changes. By stepping voltage up and down in small increments separately from frequency changes, the processor is able to reduce periods of system unavailability (which occur during frequency change). Thus, the system is able to transition between voltage and frequency states more often, providing improved power/performance balance.
  
  
• Clock Partitioning and Recovery. The bus clock continues running during state transition, even when the core clock and Phase-Locked Loop are stopped, which allows logic to remain active. The core clock is also able to restart far more quickly under Enhanced Intel SpeedStep Technology than under previous architectures.

Because Enhanced Intel SpeedStep Technology reduces the latency associated with changing the voltage/frequency pair (referred to as P-state), those transitions can be practically undertaken more often, which enables more-granular demand-based switching and the optimization of the power/performance balance based on demand. This article gives developers an overview of the support for Enhanced Intel SpeedStep Technology and demand-based switching under Linux. It is also a ready reference for developers interested in new user-level or in-kernel policy based on Enhanced Intel SpeedStep Technology.

Kernel Cpufreq Infrastructure

Cpufreq is the subsystem of the Linux kernel that allows clock speed to be explicitly set on mobile processors. A lot of recent work has advanced the modularization of generic kernel cpufreq architecture. The following figure depicts the 2.6.8 kernel cpufreq infrastructure at a high level:

Figure 1. High-level overview of the cpufreq infrastructure

The primary components of this infrastructure are as follows:

Cpufreq module: The cpufreq module provides a common interface to the various low level, CPU-specific frequency-control technologies and high-level CPU frequency-controlling policies. cpufreq decouples the CPU frequency-controlling mechanisms and policies and helps in independent development of the two. It also provides some standard interfaces to the user, with which the user can choose the policy governor and set parameters for that particular policy governor.

CPU-specific drivers: Various low-level, CPU-specific drivers implement various CPU frequency-changing technologies, such as Intel SpeedStep Technology, Enhanced Intel SpeedStep Technology, and Pentium® 4 processor clock modulation. On a given platform, one or more frequency-modulati on technologies can be supported, and a proper driver must be loaded for the platform to perform efficient frequency changes. The cpufreq infrastructure allows the user to use one CPU-specific driver per platform. The low-level CPU-specific drivers acpi and speedstep-centrino handle Enhanced Intel SpeedStep Technology-enabled CPUs.

In-kernel governors: The cpufreq infrastructure allows for frequency-changing policy governors, which can change the CPU frequency based on different criteria such as CPU usage. The cpufreq infrastructure can show various governors that are available for use on the system and allows the user to select one governor to manage CPU frequency.

Three governors can be default boot-time governors:

• Performance governor: This governor keeps the CPU at the highest possible frequency within a user-specified range.
• Powersave governor: This governor keeps the CPU at the lowest possible frequency within a user-specified range.
• Userspace governor: This governor exports the available frequency information to the user level (through the /sys file system) and permits user-space control of the CPU frequency. All user-space dynamic CPU-frequency governors, which run as daemons and control the CPU speed, use this governor as their basis.

A new in-kernel governor called the ondemand governor, which is discussed in more detail below, is added to make the most of low-latency CPU frequency/voltage-pair transitions supported in Enhanced Intel SpeedStep Technology.

Cpufreq Kernel Interface

Cpufreq exposes a number of standard interfaces in the /sys filesystem, which the user can use to change characteristics such as the CPU frequency and the policy governor. Each interface provided by cpufreq is explained below. Note that /sys should be mounted automatically.

mkdir /sys
mount -t sysfs sys /sys 

linux: # pwd
/sys/devices/system/cpu 
linux: # ls -l
total 0
drwxr-xr-x  6 root root 0 Jul 31 22:50 .
drwxr-xr-x  8 root root 0 Jul 31 22:50 ..
drwxr-xr-x  3 root root 0 Jul 31 22:50 cpu0
drwxr-xr-x  3 root root 0 Jul 31 22:50 cpu1
drwxr-xr-x  3 root root 0 Jul 31 22:50 cpu2
drwxr-xr-x  3 root root 0 Jul 31 22:50 cpu3

linux: # pwd
/sys/devices/system/cpu/cpu1/cpufreq
linux: # ls -l
total 0
drwxr-xr-x  2 root root    0 Jul 31 22:51 .
drwxr-xr-x  3 root root    0 Jul 31 22:50 ..
r--r--r--  1 root root 4096 Jul 31 22:51 cpuinfo_max_freq
r--r--r--  1 root root 4096 Jul 31 22:51 cpuinfo_min_freq
r--r--r--  1 root root 4096 Jul 31 22:51 scaling_available_frequencies
r--r--r-
-  1 root root 4096 Jul 31 22:51 scaling_available_governors
r--r--r--  1 root root 4096 Jul 31 22:51 scaling_cur_freq
r--r--r--  1 root root 4096 Jul 31 22:51 scaling_driver
rw-r--r--  1 root root 4096 Jul 31 22:51 scaling_governor
rw-r--r--  1 root root 4096 Jul 31 22:51 scaling_max_freq
rw-r--r--  1 root root 4096 Jul 31 22:51 scaling_min_freq

linux: # cat scaling_driver 
acpi-cpufreq

linux: # cat scaling_available_frequencies 
3400000 2800000

linux: # cat scaling_cur_freq 
3400000

linux: # cat scaling_available_governors 
ondemand userspace performance

linux: # cat scaling_governor 
performance
linux: # echo “userspace” > scaling_governor 
linux: # cat scaling_governor 
userspace

linux: # cat scaling_max_freq 
3400000
linux: # cat scaling_min_freq 
2800000

Using and Tuning the Ondemand In-Kernel Governor

One of the major advantages that Enhanced Intel SpeedStep Technology brings is lower latency associated with P-state changes – on the order of 10mS. In order to reap maximum benefit, the OS must perform more-frequent P-state transitions to match the current processor utilization. Doing frequent transitions with a user-level daemon will involve more kernel-to-user transitions, as well as a substantial amount of kernel-to-user data transfer. An in-kernel P-state governor, which dynamically monitors the CPU usage and makes P-state decisions based on that information, is therefore better-suited to taking full advantage of Enhanced Intel SpeedStep Technology. The ondemand policy governor is one such in-kernel P-state governor. The basic algorithm employed with the ondemand governor is as follows:

Every X milliseconds
    Get the current CPU utilization
If the utilization is more than UP_THRESHOLD %
    Increase the P-state to the maximum frequency
Every Y milliseconds
    Get the current CPU utilization
If the utilization is less than DOWN_THRESHOLD %
    Decrease P-state to next available lower frequency

The ondemand governor, when supported by the kernel, will be listed in the /sys interface under scaling_available_governors. Users can start using the ondemand governor as the P-state policy governor by writing onto scaling_governor:

linux: # cat scaling_available_governors 
ondemand userspace performance 
linux: # echo “ondemand” >
scaling_governor 
linux: # cat scaling_governor 
ondemand

Note that this sequence must be repeated on all the CPUs present in the system. Once this is done, the ondemand governor will take care of adjusting the CPU frequency automatically, based on the current CPU usage. CPU usage is based on the idle_ticks statistics.

Because a single policy governor cannot satisfy all of the needs of applications in various usage scenarios, the ondemand governor supports a number of tuning parameters. The following explanation reveals what each tuning parameter means, as well as the default value of each:

linux: # echo “ondemand” > scaling_governor 
linux: # ls -l
total 0
drwxr-xr-x  3 root root    0 Aug  5 04:23 .
drwxr-xr-x  3 root root    0 Jul 31 22:50 ..
r--r--r--  1 root root 4096 Jul 31 22:51 cpuinfo_max_freq
r--r--r--  1 root root 4096 Jul 31 22:51 cpuinfo_min_freq 
drwxr-xr-x  2 root root    0 Aug  5 04:23 ondemand
r--r--r--  1 root root 4096 Jul 31 22:51 scaling_available_frequencies
r--r--r--  1 root root 4096 Jul 31 22:51 scaling_available_governors
r--r--r--  1 root root 4096 Jul 31 22:51 scaling_cur_freq
r--r--r--  1 root root 4096 Jul 31 22:51 scaling_driver
rw-r--r--  1 root root    0 Aug  5 04:23 scaling_governor
rw-r--r--  1 root root 4096 Jul 31 22:51 scaling_max_freq
rw-r--r--  1 root root 4096 Jul 31 22:51 scaling_min_freq

Note the new ondemand sub-directory.

linux: # cd ondemand/
linux: # ls -l
total 0
drwxr-xr-x  2 root root    0 Aug  5 04:23 .
drwxr-xr-x  3 root root    0 Aug  5 04:23 ..
rw-r--r--  1 root root 4096 Aug  5 04:23 down_threshold
rw-r--r--  1 root root 4096 Aug  5 04:23 sampling_down_factor
rw-r--r--  1 root root 4096 Aug  5 04:23 sampling_rate
r--r--r--  1 root root 4096 Aug  5 04:23 sampling_rate_max
r--r--r--  1 root root 4096 Aug  5 04:23 sampling_rate_min
rw-r--r--  1 root root 4096 Aug  5 04:23 up_threshold

This directory contains all the ondemand algorithm-tuning v ariables.

linux: # cat sampling_rate_max 
55000000
linux: # cat sampling_rate_min
55000

These times are measured in microseconds, denoting the minimum and maximum sampling rate. These values are read-only, and predetermined by the kernel as a factor of P-state transition latency.

linux: # cat sampling_rate
110000

This read-write field denotes a value in microseconds. The ondemand governor checks CPU utilization and tries to increase the CPU frequency at this rate.

linux: # cat sampling_down_factor 
10

This is a read-write integer value that is multiplied with sampling_rate and used to the rate at which the ondemand governor checks CPU utilization and tries to lower the CPU frequency.

linux: # cat up_threshold 
80
linux: # cat down_threshold 
20

These are read-write fields that stand for CPU-utilization thresholds. Whenever the current utilization is more than up_threshold, the ondemand governor will increase the frequency to the maximum. Whenever the CPU utilization is less than the down_threshold, the ondemand governor tries to reduce the frequency to the next available lower P-state.

The Userspace Govenor

The Userspace governor allows any user-space utility to change the CPU frequency, by writing some value a file under /sys.

linux: # echo “userspace” >
scaling_governor 
linux: # cat scaling_governor 
userspace
linux: # ls -l	
total 0
drwxr-xr-x  2 root root    0 Jul 31 22:51 .
drwxr-xr-x  3 root root    0 Jul 31 22:50 ..
r--r--r--  1 root root 4096 Jul 31 22:51 cpuinfo_max_freq
r--r--r--  1 root root 4096 Jul 31 22:51 cpuinfo_min_freq
r--r--r--  1 root root 4096 Jul 31 22:51 scaling_available_frequencies
r--r--r--  1 root root 4096 Jul 31 22:51 scaling_available_governors
r--r-r--   1 root root 4096 Jul 31 22:51 scaling_cur_freq
r--r-r--   1 root root 4096 Jul 31 22:51 scaling_driver
rw-r--r--  1 root root    0 Jul 31 22:51 scaling_governor
rw-r--r--  1 root root 4096 Jul 31 22:51 scaling_max_freq
rw-r--r--  1 root root 4096 Jul 31 22:51 scaling_min_freq
rw-r--r--  1 root root 4096 Aug  4 18:46 scaling_setspeed

Note a new file here, scaling_setspeed:

linux: # cat scaling_setspeed 
3400000
linux: # echo 2800000 >
scaling_setspeed 
linux: # cat scaling_setspeed 
2800000

The file scaling_setspeed is read-write. When read, it denotes the current CPU frequency. The user can write a value to this field (using the list in scaling_available_frequencies), and the CPU will change the frequency to the one specified by the user. There are different CPU frequency governors, implemented as user-space daemons. They control the frequency by reading and writing values to and from scaling_setspeed. Two such user-level daemons that are widely used are powersaved and cpuspeed.

The relative advantages of in-kernel governors and user-space governors are listed here. This will help users (or administrators) in determining what kind of governor one should use under various system-usage scenarios. The advantages of in-kernel governors include the following:

• Less overhead of mode switches and data copies.
• Useful with lower transition latencies.
• Supports more-complicated policy governors, which can use other information available in the kernel, like scheduler, interrupt load, etc.
• Independent of the OSDs and their affinity towards various user-level daemons.

The advantages of user-space governors include the following:

• Easy to implement and test.
• Easy to add new governors on the fly.
• Easy to use; should not change with kernel versions.

CONFIGuring the Kernel

Cpufreq and Enhanced Intel SpeedStep Technology-related CONFIG options look like the following:

General config menu ->
    Power management options ->
        ACPI Support

<*> ACPI Support
    [*] Processor


General config menu ->
    Power management options ->
        CPU Frequency scaling

[*] CPU Frequency scaling
    <*>   /proc/cpufreq interface (deprecated)
     [*]   /proc/sys/cpu/ interface (2.4. / OLD)

    Default CPUFreq governor (userspace)  --->
        <*>   'performance' governor
        <*>   'powersave' governor
        ---   'userspace' governor for userspace frequency scaling
        <*>   'ondemand' cpufreq policy governor

        <*>   CPU frequency table helpers
            ---   CPUFreq processor drivers
            <*> ACPI Processor P-States driver
                 [ ]   /proc/acpi/processor/../performance interface 
                     (deprecated)
            <*> Intel Enhanced SpeedStep
                 [*] Use ACPI tables to decode valid frequency/voltage pairs 
                     (EXPERIMENTAL)

            < > Intel Speedstep on ICH-M chipsets (ioport interface)
            < > Intel SpeedStep on 440BX/ZX/MX chipsets (SMI interface)
            < > Intel Pentium 4 clock modulation
                :        :        :
                :        :        :
                :        :        :

The options colored in blue are the required options to support Enhanced Intel SpeedStep Technology, cpufreq interfaces, and demand-based switching. If they are configured as modules and added after boot, care should be taken to maintain the following ordering:

Conclusion

Developers who target Linux operating systems for mobile applications should use the cpufreq module to implement Enhanced Intel SpeedStep Technology. The very low latency associated with changes to the voltage/frequency pair under this technology enables high-performance demand-based switching, which can save substantially on battery life in systems based on Intel® Centrino® mobile technology.

The programming interfaces associated with the cpufreq module provide a robust environment for the implementation of this technology. Assuming that Linux continues to gain ground as a mobile operating system, power savings in Linux-based mobile applications will become more valuable as a means of gaining competitive advantage for ISVs. Implementing good support for Enhanced Intel SpeedStep Technology today is therefore a sound strategy toward positioning mobile applications for success.

Related Links

• Intel Linux Developer Center provides white papers, case studies, and other technical resources related to the implementation and support of Linux.
  
  
• ACPI _PDC Support: Intel® Processor Driver Capabilities Bit Definitions provides an overview of Intel-specific definitions of _PDC bits for use by the BIOS and OSPM.
  
  
• Cpuspeed Project Web page* is a privately maintained site that contains substantial information about the cpufreq subsystem and its use.
  
  
• IA-32 Intel® Architecture Software Developer's Manual, System Programming Guide is an authoritative reference about the programming environment for 32-bit Intel® processors.

Note: Powersaved is a user-level CPU frequency manager found in SuSE Linux installations.

Appendix

Get the possible governors on this system:

#!/bin/bash
get_governors_on_cpu() {
    if [ “$#” -ne 1 ]
    then
echo $”Usage: get_governors_on_cpu {cpu}”
return 1
    else
cpu=$1
    fi
    dir=/sys/devices/system/cpu/cpu${cpu}/cpufreq
    if [ -d $dir ]
    then
        cd   $dir
        cat scaling_available_governors
    else
     echo “CPU” ${cpu} “ available policy get FAILED”
     return 1
    fi
    return 0
}



get_governors() {
if [ “$#” -ne 0 ]
then
    echo $”Usage: get_governors”
    return 0
fi

p1=‘pwd‘
if [ ! -d /sys ]
then
    mkdir /sys
fi
if [ ! -d /sys/devices ]
then
    mount -tsysfs sysfs /sys
fi
get_governors_on_cpu 0
RETVAL=$?
if [ $RETVAL -ne 0 ]
then
    cd $p1
    return $RETVAL;
fi

cd $p1
return 0

}

get_governors $1
exit $?

Set a policy governor on all CPUs:

#!/bin/bash
set_governor_on_cpu() {
    if [ “$#” -ne 2 ]
    then
        echo $”Usage: set_governor_on_cpu {cpu} {governor}”
        return 1
    else
        cpu=$1
        governor=$2
    fi
    dir=/sys/devices/system/cpu/cpu${cpu}/cpufreq
    if [ -d $dir ]
    then
        cd   $dir
        echo $governor > scaling_governor
        if [ ‘cat scaling_governor‘ == $governor ]
        then
            echo “CPU” ${cpu} “policy set to” $governor “governor”
            echo -n “maximum frequency:”
            cat scaling_max_freq
            echo -n “minimum frequency:”
            cat scaling_min_freq
        else
            echo “CPU” ${cpu} “policy setting to” $governor “governor FAILED”
            return 1
        fi
    else
        echo “CPU” ${cpu} “policy setting to” $governor “governor FAILED”
        return 1
        fi
        return 0
}



set_governor() {
if [ “$#” -ne 1 ]
then
    echo $”Usage: set_governor {governor}”
    echo $”Example: set_governor ondemand”
    return 0
else
    governor=$1
fi
p1=‘pwd‘
if [ ! -d /sys ]
then
    mkdir /sys
fi
if [ ! -d /sys/devices ]
then
    mount -tsysfs sysfs /sys
fi
num_cpus=‘cat /proc/cpuinfo| grep “^processor”| wc -l ‘
i=0
while [ $i -lt $num_cpus ]
do
    set_governor_on_cpu $i $governor
    RETVAL=$?
    if [ $RETVAL -ne 0 ]
    then
        cd $p1
        return $RETVAL;
    fi
    i=‘expr $i + 1‘
done

cd $p1
return 0

}

set_governor $1
exit $?

Print (periodically) the speed of all CPUs:

#!/bin/bash
get_speed_on_cpu() {
    if [ “$#” -ne 1 ]
    then
        echo $”Usage: get_speed_on_cpu {cpu}”
        return 1
    else
        cpu=$1
    fi
    dir=/sys/devices/system/cpu/cpu${cpu}/cpufreq
    if [ -d $dir ]
    then
        SPEED=‘cat /${dir}/scaling_cur_freq‘
        echo “CPU”${cpu} “:” $SPEED “kHz”
    else
        echo “CPU” ${cpu} “ current frequency get FAILED&
rdquo;
        return 1
    fi
    return 0
}



get_speed() {
if [ “$#” -gt 1 ]
then
    echo $”Usage: get_speed [time in usec(default print once and exit)]”
    return 0
fi
ONESHOT=0
if [ “$#” -eq 1 ]
then
TIME=$1
else
ONESHOT=1
fi
if [ ! -d /sys ]
then
    mkdir /sys
fi
if [ ! -d /sys/devices ]
then
    mount -tsysfs sysfs /sys
fi
num_cpus=‘cat /proc/cpuinfo| grep “^processor”| wc -l ‘
while :
do
    i=0
    while [ $i -lt $num_cpus ]
    do
        get_speed_on_cpu $i
        RETVAL=$?
        if [ $RETVAL -ne 0 ]
        then
            return $RETVAL;
        fi
        i=‘expr $i + 1‘
    done
    if [ $ONESHOT -eq 1 ]
    then
        return 0
    fi
    UTIME=‘expr $TIME*1000|bc‘
    usleep $UTIME
    echo
done
return 0
}

get_speed $1 $2
exit $?